OpenAI posts Lean formalizations for claimed Kakeya advance and ETH refutation

OpenAI released manuscripts and machine-checked proof artifacts for two major claims, while stressing that outside review has only begun and the underlying model remains unavailable.

2 min read

A blackboard filled with mathematical formulas and symbols, illustrating AI-assisted mathematical research.

OpenAI has released manuscripts and Lean proof artifacts for two unusually large claims produced with an unnamed internal model: a harmonic-analysis argument it says advances the Kakeya problem, and a deterministic subexponential-time algorithm for fixed-width SAT that would refute the Exponential Time Hypothesis if correct.

The October 26 announcement presents the material as raw research data, not settled results. OpenAI says outside mathematicians and computer scientists have begun reviewing the work. The local-smoothing package maps its Lean formalization to the accompanying manuscript and identifies consequences for Kakeya- and Stein-type conjectures. The SAT package similarly documents the formalized definitions and theorem statements behind the claimed algorithm.

A blackboard filled with mathematical formulas and symbols, illustrating AI-assisted mathematical research.

Lean checking is meaningful evidence that the encoded statements follow from the imported definitions and lemmas. It does not by itself establish novelty, confirm that every natural-language interpretation matches the formal statement, or replace independent expert review. OpenAI also describes a third result in quantum field theory and Yang–Mills theory, but says those artifacts are being withheld while specialists prepare a human-led exposition.

The canonical openai/math repository is maintained by OpenAI and licensed under Apache-2.0. It showed about 15,600 stars, 1,800 forks and seven commits when checked on October 27, 2026. Those are total counts, not measured recent growth: no reliable earlier snapshot was available, so the repository’s popularity should not be presented as a growth trend.

The underlying research model remains proprietary and unnamed. OpenAI has not released weights, an API, pricing or reproducible inference instructions, leaving the public artifacts—not the generating system—as the testable contribution for now.

Sources: OpenAI announcement, local-smoothing manuscript, SAT manuscript, repository.

mathematicsfrontier modelsOpenAIformal verification